Re: [libvirt] "Choose alarm timer" feature for kvm

On Sun, Aug 24, 2008 at 08:20:05PM +0100, Daniel P. Berrange wrote: > On Sun, Aug 24, 2008 at 08:40:46PM +0200, Tiziano M?ller wrote: > > Hi everyone > > > > In KVM it is possible to select the alarm timer being used (dynticks, > > hpet, rtc, unix). > > What exactly do these options do ? Is there anything describing the > pros/cons of the different options & is there a way to determine > what the best option is for a particular VM ? I've had a long look at the source code in KVM-73 associated with these timers and have some comments: - QEMU tests the available timer sources in order and chooses the first one which works (allegedly the "best" one). The order, on Linux, is: Timer Implementation method --------------------------------------------------------------- (1) dynticks POSIX timer_create (2) hpet Linux-specific hardware (/dev/hpet) - see below (3) rtc old /dev/rtc, millisecond accuracy (4) "unix" POSIX setitimer - Dynticks is at the top of the list, and has an extra 'rearm' operation (the others only support starting and stopping timers). But I don't understand how this rearm operation is an improvement. - HPET is second in the list and most accurate. It's backed by hardware and has sub-microsecond accuracy. - It's not clear if dynticks (== timer_create) uses HPET internally in the kernel? - I can't see how any of this would affect guests -- except that choosing a more accurate timer would be in some way 'better' because the guest would then get more accurate emulation events. > Basically I'm wondering where/if this should be exposed in the XML > format, or whether libvirt could/should just internally pick the > best one ? I can't see how libvirt can do any better than qemu/kvm is doing already. So unless someone comes along with an argument that using (eg) hpet is better for Windows guests or something, then I suggest that libvirt doesn't need to do anything.